Merlin: Multi-View Representation Learning for Robust Multivariate Time Series Forecasting with Unfixed Missing Rates

ICLR 2025 Conference Submission268 Authors

Keywords: Multivariate time series forecasting with sparse observations, Multi-View Representation Learning, Offline knowledge distillation, Multi-view contrastive learning

TL;DR: This paper proposes Multi-View Representation Learning for Robust Multivariate Time Series Forecasting with Unfixed Missing Rates

Abstract:

Multivariate Time Series Forecasting (MTSF) aims to predict the future values of multiple interrelated time series and support decision-making. While deep learning models have attracted much attention in MTSF for their powerful spatial-temporal encoding capabilities, they frequently encounter the challenge of missing data resulting from numerous malfunctioning data collectors in practice.
In this case, existing models only rely on sparse observation, making it difficult to fully mine the semantics of MTS, which leads to a decline in their forecasting performance. Furthermore, the unfixed missing rates across different samples in reality pose robustness challenges. To address these issues, we propose Multi-View Representation Learning (Merlin) based on offline knowledge distillation and multi-view contrastive learning, which aims to help existing models achieve semantic alignment between sparse observations with different missing rates and complete observations, and enhance their robustness. On the one hand, we introduce offline knowledge distillation where a teacher model guides a student model in learning how to mine semantics from sparse observations similar to those obtainable from complete observations. On the other hand, we construct positive and negative data pairs using sparse observations with different missing rates. Then, we use multi-view contrastive learning to help the student model align semantics across sparse observations with different missing rates, thereby further enhancing its robustness. In this way, Merlin can fully enhance the robustness of existing forecasting models to MTS with unfixed missing rates and achieves high-precision MTSF with sparse observations.
Experiments on four real-world datasets validate our motivation and demonstrate the superiority and practicability of Merlin.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 268